S3 Finally Drops Global Naming and Agents Hit the Complexity Wall

Top-Line Summary

S3 bucket names no longer need to be globally unique — account-regional namespaces are here after twenty years. Brett shares real numbers on running OpenBrain on AWS ($30/month, with Aurora and a fck-nat instance doing the heavy lifting), and an O’Reilly piece on “The Mythical Agent-Month” lands squarely on the tension between agent productivity and the complexity jungles they create.

Show Video

The “Pre-Show” Context

Brett’s nephew was supposed to watch Minecraft Live with him, but they got distracted playing Minecraft and missed it — though a new sulfur biome with a friendly sulfur cube mob is apparently incoming. The real pre-show, though, is a long riff on personalized software: Brett’s replacing SaaS tools with Claude-built alternatives after realizing the infrastructure cost is roughly a buck a month versus $30-50 for tools he barely uses. Travers nails the counterweight: “It’s the cost of your attention and time.”

The Engineering Rundown

• AWS MCP Server Preview — Enhanced Monitoring (12:00)

  AWS added CloudWatch metrics to the MCP server preview — invocation counts, success rates, errors, throttling. Classic observability play, and Brett points out it immediately inspired him to build the same thing for OpenBrain. He fed the article to Claude (after manually copy-pasting it because Claude can’t scrape AWS news pages), and within ten minutes had a CloudWatch dashboard showing which agents and MCPs are hitting OpenBrain and how often. The agent SOPs also got a natural language upgrade, though Brett’s still reluctant to let agents loose in any account that matters. Both hosts agree: as agents get more autonomy, observability isn’t optional.

• Amazon Bedrock — GLM 5 and Minimax M2.5 (19:00)

  Two new models on Bedrock, roughly a month after their public release. Brett frames it as classic AWS strategy — be the platform with all the models, all the primitives, and let customers pick. The practical angle: Minimax is about 0.25x credit usage in Kiro compared to Opus, so if you’re doing summarization or lightweight tasks, there’s no reason to burn expensive tokens. Brett’s considering swapping some of OpenBrain’s Haiku-based summarization to test these cheaper models against each other.

• S3 Account Regional Namespaces (22:00)

  After twenty years, S3 bucket names no longer need to be globally unique. Travers walks through the change: buckets now live in an account-regional namespace, so the format becomes bucket-name plus account number plus region. Brett’s immediate reaction is mock outrage — his go-to training demo of trying to create “my-bucket” to demonstrate global uniqueness is now dead. You can also lock down IAM policies to the namespace, which is a nice security bonus. It’s just available — no feature flags to enable.

• OpenBrain AWS Costs (26:00)

  Brett shares real numbers: OpenBrain runs about $1-2/day, with roughly 80% of that being Aurora Serverless. The first mistake was Claude setting the ACU minimum to 0.5 instead of zero, so Aurora never actually scaled down for the first few days. After fixing that, the next cost surprise was needing a NAT gateway for Lambda functions making outbound calls from the VPC — $35/month for a single NAT gateway. The fix: fck-nat on a t4g.nano, dropping that to about 40 cents a day. All-in, he’s projecting around $30/month, which Travers notes is “like a fancy Minecraft server.” On a day of heavy usage, OpenBrain saw 7,000 MCP invocations — it’s not idle infrastructure.

• Malicious LiteLLM — Supply Chain Attack (show notes)

  LiteLLM versions 1.82.7 and 1.82.8 on PyPI had a credential stealer baked in — SSH keys, AWS credentials, Kubernetes configs. The attack vector was Trivy, a security scanning tool in LiteLLM’s own CI/CD pipeline. A security tool was the entry point for the compromise. If you installed those versions, rotate everything.

• OpenAI Shutting Down Sora / Disney Pulls $1B (33:00)

  Sora is dead and Disney walked away from a billion-dollar investment. Brett never understood the commercialization path — friends told him the content controls were trivially bypassed, and the IP infringement risk was enormous. Travers frames it as “OpenAI’s side quest sickness,” noting they spent cycles on Sora, a browser project that went nowhere, and other tangents while Anthropic focused on code tooling. A reality check for anyone assuming AI video generation was about to be a product.

• OpenAI Acquiring Astral (35:00)

  OpenAI bought Astral — the company behind uv, ruff, and ty — and the team is joining the Codex division. Same playbook as Anthropic acquiring Bun in December. Brett’s take: AI companies are buying developer tools not for the products but for the talent and ecosystem lock-in, similar to how developers get IDE lock-in over time. Travers is cautiously optimistic that the acquisition keeps the projects alive rather than letting them rot in a foundation, but both acknowledge the trend is worth watching — critical open-source infrastructure owned by competing AI companies is a new dynamic.

• Codex Sandboxes (40:00)

  Brett’s been using Codex to run scheduled repo scans — checking commits from the last 24 hours for security issues and bugs. It works well for that, but the sandbox isolation means he can’t get Codex to write to OpenBrain or connect to his GitLab server. The tooling friction is real: you have to rethink how you expose functionality to a sandboxed agent, which is a fundamentally different workflow than Claude Code’s local-first approach.

• Claude Code Auto Mode (43:00)

  Auto Mode sits between manual approval and the terrifying --dangerously-skip-permissions. A classifier evaluates each action — safe stuff runs, risky stuff gets blocked or escalated. Brett’s honest take: it still makes him nervous. The dropped-production-database story from a few weeks ago looms large. What if Auto Mode decides something is safe that you’d consider risky? Travers’s answer: you need a dev machine or burner environment. The broader point lands on backups — if an agent with enough permissions decided to nuke your Git repo, and that repo was your only backup, you’re done.

• Claude Code Channels (46:00)

  Channels let you push messages into a running Claude Code session from Telegram, Discord, or iMessage — essentially texting your agent from your phone. Brett sees it as Anthropic’s answer to Open Claw’s remote-access model. The webhook receiver angle is interesting (CI results or deploy alerts landing in an active session), but Brett flags the obvious risk: if you’re out having a real conversation and a thought triggers a “hey Claude, go build this,” you might end up with features you didn’t actually need. Still in research preview.

• DeepSeek Engram (50:00)

  Engram staples a massive lookup table to an LLM — O(1) lookups into pre-computed n-gram embeddings instead of the model remembering everything through weights. At 27B parameters it outperforms MoE baselines under the same compute budget. Travers’s reaction: “That’s not the most complicated thing to think about — what if we just memorize things?” — which makes the performance gains even more striking. If the approach is this simple and this effective, there’s probably a lot more to discover in this space.

• Mamba-3 (52:00)

  Mamba-3 pushes state space models forward — at 1.5B parameters, it matches Mamba-2’s perplexity with half the state size. The MIMO design adds 1.2 points of accuracy without extra decode latency. Travers connects it to Dr. Bower’s presentation from last week about smaller, more capable models running on edge devices. Brett’s more interested in where you’d deploy something this small — edge servers, mobile devices, dedicated hardware.

• Embedded LLMs in MCP Servers (55:00)

  Inspired by last week’s conversation with David, Brett can’t stop thinking about embedding a small LLM inside an MCP server to evaluate whether incoming tool calls are appropriate. He doesn’t have a concrete use case for OpenBrain yet, but the security angle is clear — a fine-tuned watchdog model checking for prompt injection or out-of-bounds requests. Travers suggests fine-tuning against known jailbreak datasets. A cheaper Bedrock model like Minimax could handle the evaluation without blowing the budget.

• The Mythical Agent-Month (58:00)

  A riff on Brooks’s classic. The core argument: agents eliminate the easy stuff (boilerplate, refactoring, tests) but make the hard stuff harder — design, scoping, knowing what not to build. Brett confesses Kiro generated over 400 tests for his Kanban board project, and he’s now using Claude to unravel the over-engineering. The article’s key line — once codebases exceed 100K lines, agents start “chasing their tails” — resonates with both hosts. The scarcest resource isn’t typing speed, it’s taste and judgment. You have to be the one saying no.

• Better, Faster, and (Even) More (1:02:00)

  The cost to go from “random thought” to “working something” has never been lower, but speed only matters if you have systems to sustain it. The author uses CLAUDE.md and WORKLOG.md in every project — basically the same pattern Brett built with OpenBrain. The meta-insight: the people getting the most out of AI tools aren’t prompting better, they’re building infrastructure around the tools. Dotfiles, reusable scripts, standardized layouts. Travers raises the organizational question — how do you keep these tribal tools available when the person who built them leaves?

• State of the Models (1:07:00)

  Travers’s current stack: GPT 5.4 for general use, Gemini Flash for one-shot multimodal tasks (images, documents), Opus in Claude Code. Minimax M2.7 and MIMO V2 Pro are hitting the sweet spot on price-to-performance — roughly Opus/Sonnet capability at a fraction of the cost. Brett’s heads-down on Claude but using Codex 5.3 for repo scans and Kiro for answering specific AWS questions (though he still can’t get Kiro to write solutions the way Claude does).

Off-the-Clock Recommendations

• Project Hail Mary (2026 film) — Travers saw it in Ultra AVX and highly recommends it. Brett’s planning a trip to the Mississauga IMAX. The book by Andy Weir is also worth reading.
• Vibe Coding by Gene Kim and Steve Yegge — Brett’s halfway through and recommending it to his training classes. It’s making him want to explore Gastown more seriously.